Saved in:
| Main Authors: | , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2509.20074 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866915511407738880 |
|---|---|
| author | Katrapati, Ganesh Shrivastava, Manish |
| author_facet | Katrapati, Ganesh Shrivastava, Manish |
| contents | Sequence to Sequence models struggle at compositionality and systematic generalisation even while they excel at many other tasks. We attribute this limitation to their failure to internalise constructions conventionalised form meaning pairings that license productive recombination. Building on these insights, we introduce an unsupervised procedure for mining pseudo-constructions: variable-slot templates automatically extracted from training data. When applied to the SCAN dataset, our method yields large gains out-of-distribution splits: accuracy rises to 47.8 %on ADD JUMP and to 20.3% on AROUND RIGHT without any architectural changes or additional supervision. The model also attains competitive performance with? 40% of the original training data, demonstrating strong data efAciency. Our findings highlight the promise of construction-aware preprocessing as an alternative to heavy architectural or training-regime interventions. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2509_20074 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Can Constructions "SCAN" Compositionality ? Katrapati, Ganesh Shrivastava, Manish Computation and Language Sequence to Sequence models struggle at compositionality and systematic generalisation even while they excel at many other tasks. We attribute this limitation to their failure to internalise constructions conventionalised form meaning pairings that license productive recombination. Building on these insights, we introduce an unsupervised procedure for mining pseudo-constructions: variable-slot templates automatically extracted from training data. When applied to the SCAN dataset, our method yields large gains out-of-distribution splits: accuracy rises to 47.8 %on ADD JUMP and to 20.3% on AROUND RIGHT without any architectural changes or additional supervision. The model also attains competitive performance with? 40% of the original training data, demonstrating strong data efAciency. Our findings highlight the promise of construction-aware preprocessing as an alternative to heavy architectural or training-regime interventions. |
| title | Can Constructions "SCAN" Compositionality ? |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2509.20074 |